Neural networks are the most seductive tool in the trading-AI toolbox: universal function approximators that, in theory, can learn any pattern hiding in your price data. In practice, most neural-network trading projects die in one of three well-documented ways — and knowing those failure modes in advance is worth more than any architecture diagram. This article walks through how the main architectures apply to market forecasting, and the discipline that separates a research toy from a deployable model.
Which architecture for which job
The three classic ways the project dies
1. The target is nearly pure noise. The daily direction of a liquid market is close to a coin flip; the signal-to-noise ratio in financial returns is among the worst of any domain machine learning is applied to. A model with 52-53% directional accuracy can be genuinely excellent — and genuinely profitable with the right position sizing — but if you expect image-classification accuracy from price data, you will keep "improving" the model until it memorizes noise.
2. Lookahead leaks in through preprocessing. The most common silent killer is not the network — it is the pipeline. Normalizing with statistics computed over the whole dataset (including the test period), creating features with centred rolling windows, or shuffling before splitting all leak future information. The symptom is always the same: spectacular backtest, immediate live decay. Fit every scaler and every feature transformation on training data only, and split time series chronologically — walk-forward, never randomly.
3. It beats nothing. Every neural forecaster must be benchmarked against embarrassingly simple baselines: predicting yesterday's value, the unconditional mean, a linear regression, a moving-average crossover. A large share of published and hobbyist results fail this test once transaction costs enter. If the network cannot clearly beat a linear model on out-of-sample data after costs, deploy the linear model — it is faster, more stable and easier to debug.
A sane workflow
The bottom line
Neural networks absolutely have a place in trading — as one more member of a disciplined pipeline, next to honest validation and boring baselines. They are not a shortcut around statistics. Treat the 52% model with respect, guard the pipeline against leakage like a hawk, and remember that in this domain the network is rarely the hard part; the process around it is.
Which architecture for which job
- Feedforward networks (MLPs) map a fixed vector of features (returns, indicator values, volatility measures) to a prediction. They know nothing about time — the sequence is whatever you engineer into the features. Cheap to train and a surprisingly strong baseline.
- Recurrent networks, especially LSTMs, read a sequence step by step and carry a memory cell forward, which lets them capture longer dependencies without the vanishing-gradient problem of plain RNNs. For years they were the default for financial sequences.
- Convolutional networks (1D CNNs) slide filters over a window of prices, detecting local shapes — momentum bursts, volatility clusters — regardless of where they occur. Faster to train than LSTMs and often competitive.
- Attention/transformer models look at the whole sequence at once and learn which time steps matter. State of the art in language, but data-hungry — and a daily price series has thousands of observations, not billions of tokens. On small financial datasets they overfit spectacularly without heavy regularization.
The three classic ways the project dies
1. The target is nearly pure noise. The daily direction of a liquid market is close to a coin flip; the signal-to-noise ratio in financial returns is among the worst of any domain machine learning is applied to. A model with 52-53% directional accuracy can be genuinely excellent — and genuinely profitable with the right position sizing — but if you expect image-classification accuracy from price data, you will keep "improving" the model until it memorizes noise.
2. Lookahead leaks in through preprocessing. The most common silent killer is not the network — it is the pipeline. Normalizing with statistics computed over the whole dataset (including the test period), creating features with centred rolling windows, or shuffling before splitting all leak future information. The symptom is always the same: spectacular backtest, immediate live decay. Fit every scaler and every feature transformation on training data only, and split time series chronologically — walk-forward, never randomly.
3. It beats nothing. Every neural forecaster must be benchmarked against embarrassingly simple baselines: predicting yesterday's value, the unconditional mean, a linear regression, a moving-average crossover. A large share of published and hobbyist results fail this test once transaction costs enter. If the network cannot clearly beat a linear model on out-of-sample data after costs, deploy the linear model — it is faster, more stable and easier to debug.
A sane workflow
- Predict returns or direction, never raw prices — prices are non-stationary, and a network that predicts "tomorrow will be near today" scores a deceptively low error while saying nothing.
- Start with a small MLP on well-engineered features; add sequence models only if they demonstrably improve walk-forward results.
- Use dropout, early stopping and weight decay aggressively — financial data rewards regularization far more than depth.
- Retrain on a schedule and monitor live predictions against the backtest distribution: markets drift, and every model has a shelf life.
- Size positions from predicted probability, not from a binary signal — the confidence of the network carries usable information.
The bottom line
Neural networks absolutely have a place in trading — as one more member of a disciplined pipeline, next to honest validation and boring baselines. They are not a shortcut around statistics. Treat the 52% model with respect, guard the pipeline against leakage like a hawk, and remember that in this domain the network is rarely the hard part; the process around it is.